Study: LLM Reasoning Relies on Reusable Procedures Learned in Pretraining

A study by Laura Ruis et al. suggests that LLM reasoning is not driven by retrieving parametric knowledge memorized during pretraining, but by reusable procedures learned in pretraining, offering new insight into how large language models reason.

2026-09-30 ~ 2026-09-30 · 2 related posts

• LLM reasoning is learned reusable procedures, not retrieved parametric knowledge — egrefen · 2026-09-30
• Study: LLM reasoning runs on procedural knowledge learned in pretraining, not memorized answers — egrefen · 2026-09-30